Feature/add googlenet #66
Conversation
… paper and added a method for repeating the spectrogram tensor in 3 channels
# Conflicts: # requirements.txt # soundbay/conf_dict.py
…retrained weights
… last layers size mismatch
…of 3 channel spectrogram)
…of 3 channel spectrogram)
…files' into feature/add_googlenet
|
|
Z30G0D
added
ci-test
There was a problem hiding this comment.
Great PR!
Important, your test seems to be failing on one of your new imports from torchvision.
Try to consult this post:
https://stackoverflow.com/questions/70998767/no-module-named-torchvision-models-utils
Although I am concerned about this message,
Note that this syntax is only for higher versions of PyTorch.
We might need to upgrade torch for this.
| @@ -0,0 +1,8 @@ | |||
| # @package _global_ | |||
There was a problem hiding this comment.
Why does this file exist?
Should there be so many new yaml files?
There was a problem hiding this comment.
I think that different types of experiments (e.g. with many different parameters such as model, preprocessing, optimizer, and not just one different parameter) are easer to handle when they have their own main. If it's only one parameter like a dataset path it's simple enough to add an override to the train command line. The new "optim", "preprocessors" and "model" yamls are of course required.
But on a related note, I may have misunderstood the changes you made in the latest PR to the hierarchy, I thought that every "runs" conf file should now be in the parent folder in order to reach all the relevant yamls, so I moved main, main_inference and the parallel files for the googlenet manatees experiment up to the conf folder.
| wandb.log(log_specs, commit=False) | ||
| #avoid spectrogram upload if using 3 channel spectrograms (in googlenet), upload only wavs | ||
| if audio.size()[1] == 3: | ||
| log_wavs = {f'First batch {flag} original wavs': list_of_wavs_objects} |
There was a problem hiding this comment.
Need to notify user somehow about this, he will be surprise.
Maybe use a warning or info statement to notify about this.
Should appear only once or twice during the run though.
There was a problem hiding this comment.
good idea, I'll check how to do that
amitgalor18
added
ci-test
Adding googlenet model to models, with a transform to the spectrograms to make them 3 channeled (RGB) in order to use a pretrained googlenet, based on the JASA paper. Creating conf files to enable experiments with the googlenet model